[Expanded version of paper given at MT Summit VI, 1997]
FIRST STEPS IN MECHANICAL TRANSLATION
John Hutchins
(University of East Anglia, Norwich, UK)
Abstract
Although the first ideas for mechanical translation were made in the seventeenth century, it was
not until this century that means became available for realisation with the appearance of the
electronic computer in the mid 1940s. Fifty years ago, in March 1947 Warren Weaver wrote to
Norbert Wiener and met Andrew Booth, mentioning to both the use of computers for translation.
The possibilities were investigated during the next seven years, until in January 1954 the first
prototype program was demonstrated. This article is a chronicle of these early years of
mechanised translation.
1. The beginnings to 1947
The idea of mechanising translation can be traced back to the seventeenth century, almost
always in association with the idea of inventing a universal language. In the previous century
Francis Bacon had stressed the inadequacies of natural languages for the description of scientific
discoveries. By the early seventeenth century the need was more widely felt, and the rapid
growth of international trade throughout the world and the decline of knowledge of Latin made
the dream of a universal language even more desirable. As a result many proposals for 'universal
characters' were put forward; the best known being that of John Wilkins in 1668 for the newly
founded Royal Society of London. These schemes for universal characters or 'common writing'
were invariably furnished with dictionaries from a vernacular language or from Latin, which
were intended to be applied by users in a 'mechanical' fashion. It was in this spirit that Descartes
made a comment in a letter to Marin Mersenne on 20th November 1629. It may be freely
translated:
Thus the sole utility I can see arising from this invention is for writing: if a large
dictionary is printed in all the languages in which one wants to be understood, where
common characters are put for each primitive word corresponding to the sense and
not to the syllables, such as the same character for aymer, amare and filein; [so that]
those having the dictionary and knowing the grammar would be able, through
seeking all these characters one after another, to understand in their language what is
being written.
Descartes' comment can be read as a prefiguration of how an interlingual 'mechanical' dictionary
might work. In this sense, the glossaries attached to universal languages might be seen as
forerunners of 'mechanical dictionaries'. This was indeed been the case with a scheme published
in 1661 by a German chemist Johann Joachim Becher. At the height of enthusiasm for MT in the
1960s, his work was re-published as an form of pre-computer program for machine translation
(Becher 1962).
However, as the century progressed, the search for an ideal language of communication -a goal pursued by many famous philosophers, notably by Leibniz – was recognised to be
unrealistic. The search was virtually abandoned (apart from a brief revival in the decade after the
French Revolution) until the last years of the nineteenth century. It is then that there first
appeared various schemes for international languages, the best known being of course Esperanto.
However, it was not until the twentieth century that the means became available to
develop practical devices for mechanising translation. In 1933 two inventions were patented
independently. One was in France by Georges Artsrouni, the other in Russia by Petr Troyanskii.
In both cases, the patents were essentially electrical devices for mechanical dictionaries
(Hutchins 1993). Artsrouni's invention was exhibited in 1937 at the World Exhibition in Paris,
and attracted much attention. Apparently under development since 1929 Artsrouni's invention
was intended for a wide variety of tasks – the production of timetables and telephone books, for
accounting, and for deciphering and encrypting messages, as well as for translating. In essence,
the device comprised four components: a 'memory' of dictionary entries on a paper tape, a
keyboard for entering words, a search mechanism, and a means of displaying results in a series
of windows on the keyboard.
While Artsrouni's invention was no more than a mechanical dictionary, Troyanskii went
further. His devise consisted of a mechanically operated table containing a multilingual 'glossary
field' by which words of the source language were presented through an aperture and equivalent
words of a target language were printed out. The words appeared not in full forms but as stems,
and Troyanskii devised a set of 'logical symbols' common to all languages for expressing
grammatical relationships. These symbolic characters were clearly based on Esperanto. For
example, the suffix -aj indicated a plural noun, -n was used to express a direct object case, -ir a
verb form, etc. Troyanskii envisaged translation as a three-stage process. A human 'editor' would
prepare the source text by reducing words to stems and by identifying the relevant 'logical'
relationships. The prepared text would then be converted wholly mechanically by his device into
an equivalent sequence of words in a target language, while retaining the logical symbols. A
human 'editor' would then convert the string of stems and symbols into the target language.
In the years following his 1933 patent, Troyanskii worked on various technical
improvements and at some point suggested that the processes of analysis and generation might
also be mechanised. Troyanskii clearly underestimated the amount of work involved in analysing
texts into stems and logical symbols, or in generating texts from them. He was aware that
homonyms and lexical differences between languages had to be dealt with, but he seems to have
been quite unaware of the problems of syntactic restructuring. Nevertheless, there is no doubt
that had he lived to see the first computers, Troyanskii would have been one of the pioneers of
machine translation.
2. March 1947
These precursors were, however, quite unknown in the mid 1940s, when the first
electronic computers were being developed and when the first suggestions were made to apply
their formidable powers to the task of translating languages. (A fuller account of the MT pioneer
years is given in Hutchins, 1997), where details of all the sources cited here will also be found.)
We can therefore date the true beginning of machine translation (i.e. the application of
computers to translation) as the 4th of March 1947. On this day, Warren Weaver, director of the
Natural Sciences Division of the Rockefeller Foundation, wrote to the cyberneticist Norbert
Wiener:
Recognizing fully, even though necessarily vaguely, the semantic difficulties
because of multiple meanings, etc., I have wondered if it were unthinkable to design
a computer which would translate. Even if it would translate only scientific material
(where the semantic difficulties are very notably less), and even if it did produce an
inelegant (but intelligible) result, it would seem to me worth while.
Weaver, who had heard about the success of computers in cryptography, wondered
if the problem of translation could conceivably be treated as a problem in
cryptography. When I look at an article in Russian, I say 'This is really written in
English, but it has been coded in some strange symbols. I will now proceed to
decode.'
Wiener's reply was disappointing, arguing that the problems of semantics were too great for
successful mechanisation.
However, just two days after writing to Wiener, Weaver met Andrew Booth, a British
crystallographer, who was in the United States investigating the newly invented computers. He
had come to Weaver to discuss possible American funding for a British computer at the
University of London. Weaver did not think there would be any funds for numerical applications
but he thought that there might be an interest in supporting new non-numerical applications of
computers, and he suggested translation.
The idea of MT had come to Weaver some time earlier, quite possibly in 1945, and he
may well have spoken to people about it. One of these was probably J.Desmond Bernal, who was
head of Booth's department at Birkbeck College and was a regular visitor to the US and the
Rockefeller Foundation.
At a later date, Booth stated that the discussion about MT took place in June 1946 (Booth
and Locke 1955) - and this statement has been taken as accurate by many later writers. It is
certainly true that he met Weaver on June 1946 and that there was discussion of the funding of a
British computer. However, there is no documentary evidence that Weaver raised the idea of MT
on this occasion. Indeed, Booth himself stated in other papers of the time that the meeting had
been in March 1947.
3. Booth and Richens
Weaver's discussion with Booth bore fruit. On his return to England, Booth did not forget
about MT. In his report to the Rockefeller Foundation in February 1948 he included MT in a list
of possible applications for the proposed US-funded computer at the University of London:
...We have considered this problem in some detail and it transpires that a machine
of the type envisaged could perform this function without any modification in its
design.
He was probably referring to the programming of a mechanical dictionary, which he had begun
with Kathleen Britten (his future wife) during 1947 while at the Institute for Advanced Study in
Princeton. However, Booth had subsequently met Richard Richens.
Richens was Assistant Director of the Commonwealth Bureau of Plant Breeding and
Genetics in Cambridge. His interest in mechanical translation had arisen out of experiments with
punched cards for storing information:
The idea of using punched cards for automatic translation arose as a spin-off,
fuelled by my realisation as editor of an abstract journal (Plant Breeding Abstracts)
that linguists conversant with the grammar of a foreign language and ignorant of
the subject matter provided much worse translations than scientists conversant with
the subject matter but hazy about the grammar
He proposed the segmenting of words into their base forms (stems) and endings (e.g.
inflections), both to reduce the size of dictionaries and to use grammatical annotations to
augment a strictly word-by-word dictionary 'translation'.
The validity of Richens' method was tested by hand and by using punched card
machinery on a variety of languages. It was not apparently implemented as a computer program,
primarily because the authors did not have access to a computer of sufficient size and reliability.
The experiment was written up in a memorandum during 1948, although not published until
much later, in a revised version (Richens and Booth 1955). From this account, the crudity of the
output is obvious:
(French) Il n'est pas étonn*ant de constat*er que les hormone*s de croissance
ag*issent sur certain*es espèce*s, alors qu'elles sont in*opér*antes sur d'autre*s, si
l'on song*e à la grand*e spécificité des ces substance*s.
(English) v not is not/step astonish v of establish v that/which? v hormone m of
growth act m on certain m species m, then that/which? v not operate m on of other
m if v one dream/consider z to v great v specificity of those substance m.
(The asterisks in the French indicate automatic segmentations. in the English 'translation' v indicates a
French word not translated, m "multiple, plural or dual", z "unspecific", and slashes alternative
translations.)
Although, only the very first stage of a true translation system, Richens is rightly credited
as being the first person to propose a method of automatically identifying word-endings, albeit
not using a computer but a punched card machine.
4. Developments in Los Angeles
By this time there were signs of the start of research in the United States. In May 1948
the National Bureau of Standards (NBS) decided to build computers in Washington and in Los
Angeles at its Institute of Numerical Analysis. Harry D.Huskey, who had worked for a year in
England at the National Physical Laboratory (NPL), led the team building the computer which
was eventually called the Standards Western Automatic Computer (SWAC). Either
independently or perhaps from conversations with Turing at NPL, Huskey proposed to use
SWAC for translation. In May 1949, the machine was demonstrated and a report appeared in the
New York Times, the first newspaper article on MT:
A new type of "electric brain" calculating machine capable not only of
performing complex mathematical problems but even of translating foreign
languages, is under construction here... While the exact scope of the machine will
have in the translating field has not been decided, the scientists working on it say it
would be quite possible to make it encompass the 60,000 words of the Webster
Collegiate Dictionary with equivalents for each word in as many as three foreign
languages.
After describing how the computers of the time operated, the report continued:
When a foreign word for translation is fed into the machine, in the form of an
electro-mathematical symbol or a tape or card, the machine will run through its
"memory" and if it finds that symbol as record, will automatically emit a
predetermined equivalent - the English word.
The reporter recognised the limitations of "a crude word-for-word translation" but saw the
potential value as "scientists' translations of foreign technical papers". An editorial the next day
(one of the few ever devoted to MT) expressed doubts about the accuracy:
How is the machine to decide if the French word "pont" is to be translated as
"bridge" or "deck" or to know that "operation" in German means a surgical
operation? All the machine can do is to simplify the task of looking up words in a
dictionary and setting down their English equivalents on a tape, so that the
translator still has to frame the proper sentences and give the words their contextual
meaning.
Quite wisely, it commented that:
We are still far from the machine into which we will pour cards and pull out great
poetry or great novels. In fact we shall never reach that stage.
Nevertheless, the editor shared the common assumption that the awesome mathematical power
of computers made them capable of anything.
5. Weaver's memorandum, July 1949
In July 1949, two years after writing to Wiener and meeting Booth, Weaver wrote the
memorandum (Weaver 1949) which was to launch MT as a serious subject of research in the
United States, and subsequently throughout the world. His aim was to summarise the work
which had already been done and to outline the directions he thought research could take.
He begins by describing the origins of his own interest in the topic. First he recounts a
story about decoding Turkish, which had succeeded without the decipherer having any
knowledge of the original language. The success of cryptography lay, he believed, on
"frequencies of letters, letter combinations, intervals between letters and letter combinations,
letter patterns, etc., which are to some significant degree independent of the language used"; and
he supposed, therefore, that:
there are certain invariant properties which are, again not precisely but to some
statistically useful degree, common to all languages.
In support he cites the findings of logicians such as Hans Reichenbach and linguists such as
Erwin Reifler; the latter having noted that "the Chinese words for 'to shoot' and 'to dismiss' show
a remarkable phonological and graphic agreement" – a range of meaning found also in English
'fire', which Weaver thought was more than a coincidence.
After reproducing his letter to Norbert Wiener and Wiener's disappointing reply, Weaver
reports what had been done so far. He had visited England and heard about Richens' work with
Booth; and he outlines Richens' ideas on suffix-cutting. He mentions also the newspaper reports
of activity in California, although strangely without saying where it was undertaken or who was
directing the work.
There had been no actual translation by computer – the work of Booth and Richens was
simulated on punched cards only, and from California it was merely a proposal to apply a
computer still under construction. Weaver's main concern was to indicate how MT might go
beyond the limitations of word-for-word translation, and he proposed, as he puts it, "at the risk of
being foolishly naïve... four types of attack".
The first approach was to tackle multiple meanings by examination of immediate context.
How much context would be required could vary from subject to subject, and from one word to
another, and he thought an investigation would be useful. His second proposal assumed that
there are (at least in non-literary works) logical elements in language. He believed that MT could
be based on the theorem proved in 1943 by McCulloch and Pitts on logical deduction of proofs
by automata, which suggested that "insofar as written language is an expression of logical
character" the problem of translation by computer is formally solvable.
It was linked to a third proposal involving the recently developed 'information theory' of
Claude Shannon, which was concerned with the basic statistical properties of communication.
Although enthusiastic for this approach, Weaver did stress the fundamentally probabilistic nature
of language, and the inevitable imperfection of any translation:
"Perfect" translation is almost surely unattainable. Processes, which at stated
confidence levels will produce a translation which contains only X per cent "error,"
are almost certainly attainable.
But in his fourth proposal, he was more utopian. The "most promising approach of all" is to
investigate the fundamental structures of languages:
Think, by analogy, of individuals living in a series of tall closed towers, all
erected over a common foundation. When they try to communicate with one
another, they shout back and forth, each from his own closed tower. It is difficult to
make the sound penetrate even the nearest towers, and communication proceeds
very poorly indeed. But, when an individual goes down his tower, he finds himself
in a great open basement, common to all the towers. Here he establishes easy and
useful communication with the persons who have also descended from their
towers.
Thus it may be true that the way to translate from Chinese to Arabic, or from
Russian to Portuguese, is not to attempt the direct route, shouting from tower to
tower. Perhaps the way is to descend, from each language, down to the common
base of human communication – the real but as yet undiscovered universal
language – and then re-emerge by whatever particular route is convenient.
He readily admits that this approach involves a "tremendous amount of work in the logical
structures of languages before one would be ready for any mechanization", although he believes
that Ogden and Richard's work on Basic English represented some steps towards it.
6. Contemporary proposals
While we can credit Weaver with the most explicit proposals about MT during the early
years of the computer, it is quite clear that the idea of mechanised translation was an idea that
had come to others at this time.
There are a number of instances where proposals for translation by computer were put
forward independently. In 1946, Duncan Harkin had been inspired by studies of word counts to
think of a system based on the most frequent vocabulary, and in collaboration with an electronics
engineer, planned the outline of a machine for mechanical translation. In September 1947 Alan
Turing wrote a report for the National Physical Laboratory about his plans for constructing a
computer, in which he mentioned MT as a possible demonstration of the 'intelligence' of
computers. In August 1949, the New York Times reported from Salerno that an Italian named
Federico Pucci, had invented a machine to translate, saying that it would be exhibited at a Paris
Fair; but no more was to be heard of it. In February 1950 Charles C.Holt of Chicago revealed in
a letter to the Rockefeller Foundation that he had thought of using cathode ray tubes as a storage
device for a mechanical dictionary – in some respects, it was similar in concept to the
mechanism put forward by Troyanskii in the 1930s.
There may well have been others. The new electronic computers had caught the
imagination of many people. Reports on the 'electronic brains' – the term regularly used by
journalists – appeared almost daily in national newspapers throughout the world. Using them for
translation occurred to a number of people.
7. Reactions to the memorandum
Some recipients of Weaver's memorandum were inspired to take up the challenges of MT
immediately. First on the scene was the sinologist mentioned in the memorandum itself, Erwin
Reifler. Already by 10th January 1950, Reifler produced a 55 page study, in which he put
forward the first formulated conceptions of pre- and post-editing. Assuming that 'mechanical'
processes could only be word for word one-to-one substitutions, there were obvious
inadequacies. Reifler suggested that in order to remove ambiguities from the source text a human
"pre-editor" could add extra symbols for grammatical and logical explicitness. Theoretically a
pre-editor need not know the output language. The task of the post-editor was to render the
machine output into a reasonably literate form – and, theoretically too, the post-editor need not
know the source language. Reifler argued that without knowledge of the source language a posteditor would have great difficulty faced with lists of alternatives for each word; the ambiguities
of the original could be more easily resolved by the pre-editor who would know the linguistic
and cultural context. Reifler envisaged that the pre-editor would have access to a monolingual
dictionary presenting all the alternative interpretations for those words with more than one
possible translation in the target language.
Weaver's memorandum had stimulated interest also at the Rand Corporation in Santa
Monica. Following up Weaver's suggested statistical approach to resolving problems of multiple
meaning, Abraham Kaplan investigated the micro contexts of polysemy in mathematical texts.
His study was completed on 30 November 1950, concluding that the "most practical context is ...
one word on each side, increased to two if one of the context words is a particle", i.e. an article,
preposition or conjunction. Despite its limitations and the tentativeness of its conclusions, this
study was to encourage hopes that problems of ambiguity could be resolved, and that statistical
analyses could contribute useful data for MT systems.
Also in California, 1950 saw the beginning of research at UCLA, initiated by Huskey. In
the Department of Spanish, William E.Bull began his statistical studies of vocabulary, and in the
Department of German Victor A.Oswald, Jr. researched syntactic problems in collaboration with
Stuart L.Fletcher of the NBS. This was the first research in MT devoted to syntactic questions,
and it resulted in the first article on MT to be published in an academic journal (Oswald and
Fletcher 1951). The authors proposed the coding of German sentences to identify 'noun blocks'
and 'verb blocks' and to determine which blocks were candidates for rearrangement after
translation into English. For example, using a passage from a mathematics text by Cantor:
Bevor wir diese Definition im Einzelnen zergliedern, wollen wir einige Beispiele von
Mengen betrachten, die uns anschauliches Material zum Verständnis der Definition liefern
sollen.
This sentence was to be rearranged, on the basis of identified 'blocks', as:
Bevor wir zergliedern diese Definition im Einzelnen, wir wollen betrachten einige
Beispiele von Mengen, die sollen liefern uns anschauliches Material zum Verständnis der
Definition.
In this sequence, the German could be given a potential word-for-word translation into English
as:
Before we analyze this definition in detail we want-to regard some examples of sets,
which shall furnish us perceptible material for-the understanding of-the definition.
Their conclusion was that syntax "does not constitute, as had been thought by some, a barrier to
mechanical translations." it was a rather optimistic conclusion, but, we should remember, it was
not based on any test by a computer; although the procedures had been formulated as
'instructions' for the SWAC, they were not implemented.
8. Bar-Hillel's survey 1951
In May 1951, Yehoshua Bar-Hillel was appointed as a research assistant in the Research
Laboratory for Electronics at the Massachusetts Institute of Technology (MIT). His task was to
study the possibilities of MT and to plan future research, and in October 1951 he visited the few
places in the US where MT research was going on. After the tour it was decided to hold a
conference at MIT in June 1952, and in preparation for it Bar-Hillel wrote a state-of-the-art
report which represents a milestone in the development of MT.
His survey (Bar-Hillel 1951) begins by summarizing the presumed benefits of MT to
satisfy translation demands, particularly in science, finance and diplomacy; and to provide means
for scanning at high speed ("though perhaps low-accuracy") the huge printed output in
newspapers, journals, leaflets of actual or potential enemies. As a side-product MT could also
throw light on aspects of linguistics and communication.
Although he could envisage no hope of achieving "high-accuracy, fully automatic MT...
in the foreseeable future", he stressed that this did not mean computers could not be applied to
translation: "with a lowering of the target, there appear less ambitious aims the achievement of
which is still theoretically and practically viable." Therefore, Bar-Hillel reviewed various
options for "mixed MT, i.e. a translation process in which a human brain intervenes."
A post-editor was considered indispensable to eliminate "semantical ambiguities", but he
stressed the importance of not printing out all possible translations of every word in the sentence
order of the original. What was clearly needed was for the grammatical ambiguities to be
resolved automatically, and he made suggestions for an "operational syntax". He did not agree
with Oswald and Fletcher that syntactic analysis must await the availability of complete
statistical data for morphological analysis in order to overcome the limitations of computer
storage. He was already profoundly sceptical of statistical approaches -- a view which he was to
repeat in subsequent years. Statistics might well identify the most frequent words and endings
and enable 90 per cent of an average text to be translated, but the result would be unsatisfactory
since the post-editor would be faced with translating the words which are "the least predictable
and highly loaded with information".
He then briefly considered the role of a pre-editor as envisaged by Erwin Reifler. The
problem with Reifler's proposals was the construction of a monolingual dictionary which the preeditor could use to select appropriate translation equivalents in an unknown target language. BarHillel agreed that it might be possible where only two languages are concerned, but not for
"general MT, where translation from any language into any other is considered."
Whereas, in Bar-Hillel's view, "specific MT" could be pursued on a simple trial-and-error
basis, "general MT will require establishment of a universal, or at least general grammar,
perhaps even the construction of a whole artificial exchange-language." He was aware of failed
attempts in previous centuries to construct universal languages, but believed that "mathematical
logic, and modern structural linguistics" may prove a better foundation.
In the final pages of his survey, he mentions the possibility of MT of texts with "a
restricted vocabulary or a restricted number of sentence-patterns." Here he had in mind Basic
English, artificial languages such as Esperanto, and the codes used by pilots and meteorologists.
He thought it might be possible in such cases to translate whole units or sentences, and "the
theoretical difficulties... are clearly less formidable". But he went further: there was also the
"possibility of restricting, by voluntary convention, the richness of expression", i.e. by
establishing what are now known as 'controlled languages'.
Bar-Hillel's paper was written before any MT research on even a reasonably modest scale
had begun, in particular before any computer had been programmed to do even word-for-word
translation, and it is therefore remarkable that he was able to identify so clearly the main
problem areas and to outline many of the basic strategies which continue to be valid to the
present day.
9. The first MT conference, June 1952
Bar-Hillel organised the first conference on machine translation at the Massachusetts
Institute of Technology from the 17th to the 20th June 1952. Reports of this milestone event are
to be found in papers by Reynolds (1954) and Reifler (1954).
At its opening public session Bar-Hillel outlined the history of MT, from Weaver's first
ideas to the establishment of the research groups represented at the conference. After pointing
out the obvious potential of MT to cope with the immense and growing volume of materials to
be translated, Bar-Hillel went on to stress the complexities of mechanising translation processes,
concluding that
completely automatic and autonomous mechanical translation with unique correlates to the
original text is, in general, practically excluded, even with respect to scientific texts... This
being so, machine translation means no more than mechanical aids to translation. Only
some kind of brain-machine partnership is envisaged.
As in his 1951 review, Bar-Hillel mentioned the partnerships of post-editors and preeditors. He conceded that economically MT was not yet viable:
Even if it should turn out that none of the possible machine-brain partnerships
would be more effective than a human translator, in the sense that they will be
neither quicker nor cheaper nor more exact than the human translator, under
conditions existing today or in the near future, I would strongly advocate a
continuation of this research. Electronic machines will doubtless become cheaper,
human brains probably more expensive. A partnership that could not stand free
competition today may well outbid its human competitors in some not too remote
future.
All participants agreed on the need for post-editing. At the conference, Bar-Hillel repeated what
he had said on the topic in his survey. Then in another paper he argued that post-editors should
be given every possible version. This paper, which was concerned with idioms, suggested that
'idiomatic' phrases should either be included as units in the lexicon or that post-editors should be
alerted to the fact that certain output forms might be replaced by phrases. His example was Es
gibt einen Unterschied, where the correct translation should be There is a difference. Inclusion
of es gibt as a phrase in the dictionary with the automatic translation there is/are would prevent a
literal translation when required, e.g. es might refer to Mädchen, in which case she gives... would
be correct. He argued therefore that post-editors should be presented with all the alternatives. In
fact, Bar-Hillel was of the opinion that if MT is limited to Western languages closely related to
English and to scientific publications, and if the dictionary is limited to 'non-overlapping
synonyms' then multiple meanings can all be dealt with satisfactorily by a post-editor.
Many agreed with Reifler that pre-editing of some kind was also essential to minimise
ambiguities and syntactic complexities of source texts. Alternatively, it was thought that authors
could be trained to write with MT in mind, e.g. by using a 'controlled language'. At the
conference, Stuart Dodd of the Washington Public Opinion Laboratory presented his ideas for
the "standardization of English syntax as a means of simplifying the use of English either as a
source language or as a target language". This included the regularisation of verb forms (She did
be loved instead of She was loved) and of pronouns (using only nominative forms: I will send he
to she) and, of course, the use of words in one meaning only.
Reifler elaborated on the pre-editing codes. These were to be "a graphic supplementation
of the conventional form of the foreign message which raises its graphic-semantic explicitness to
the level necessary for a mechanical translation." The coding would take into account not just the
multiple meanings in the source language but also lexical and semantic differences in the target
language. He thought the codes could be inserted by writers themselves using special
monolingual dictionaries with symbols for distinguishing homographs. But he had also a more
radical proposal: a new orthography for all languages which would distinguish grammatical
categories: "all nouns would have... a capital first letter..., all principal verbs with a capital
second letter and all attributive adjectives with a capital third letter..."; so that for example, the
German er hegt die fromme Hoffnung would be written "er hEgt die frOmme Hoffnung." The
idea was that this would ease the specification of the context in which Kaplan's method of
disambiguation could operate. Furthermore, in conjunction with a regularisation of the target
'model' language, it would "either restrict post-editorial interpretation to a minimum, or it may
even make it completely superfluous."
For translation into many target languages Reifler agreed with Bar-Hillel that this raised
the question of a 'universal grammar'. He believed that comparative-historical linguistics could
help to identify real universals, but there may also be 'pseudo-universals'. He thought that words
of source languages could be assigned grammatical categories arbitrarily aligned to categories of
target languages. For example, the Mandarin Chinese version of English he walks quickly is t'a1
tsou3-ti k'uai4, where k'uai4 corresponds to something like "to be quickness" or "to be quick"
and tsou3-ti corresponds to "walk's" or "of walk". But in other contexts tsou3-ti can be translated
as "walking"; if there were an arbitrary equation of -ti and English -ing, the sentence could be
rendered as "he walk-ing quick" – a kind of pidgin English.
Although Weaver had suggested investigations of universal languages for MT, the topic
was not discussed in the conference. For translation into many languages, where now we would
expect proposals for some kind of interlingua, there was instead a suggestion from Leon Dostert
of Georgetown University that:
general MT (mechanical translation from one into many languages)... should be so
developed that one translates first from the input language into one 'pivot' language
(which in our case will, most likely, be English) and from that pivot language into
any one of the output languages desired.
The conference was attracted by the ideas of Victor Oswald for domain-specific
dictionaries ('microglossaries' he called them) to minimise the problems of "multiple meanings".
The 'sublanguage' vocabulary (as it would now be called) was to be identified by statistical
analysis of corpora on the basis of the observation that "the data of all frequency counts fall into
the same pattern, which means that a frequency count of any micro-segment of any language –
say the nouns in German contexts pertaining to brain surgery – should give a parabolic curve
where high-frequency elements ought to dispose of eighty-percent of all running nouns." Not
only was this found to be true, and that familiarity with 80% of the technical words for any
article was alone enough to make sense of the article, but Oswald found a similar frequency
distribution for the non-technical words.
William E.Bull was sceptical about the value of frequency analyses for constructing
micro-glossaries:
There exists no scientific method of establishing a limited vocabulary which will
translate any predictable percentage of the content (not the volume) of
heterogeneous material... A micro-vocabulary appears feasible only if one is
dealing with a micro-subject, a field in which the number of objective entities and
the number of possible actions are extremely limited. The number of such fields is,
probably, insignificant...
Indeed, Bull went on to pinpoint a basic problem for all MT systems to the present day:
The limitations of machine translation which we must face are, vocabularywise,
the inadequacy of a closed and rigid system operating as the medium of translation
with an ever-expanding, open continuum.
Syntax was something quite new for many participants. Victor Oswald spoke about his
analysis of German syntax: his 'blocks' approach was in essence Oswald a form of constituency
analysis, which was becoming familiar at the time. More of a revelation was Bar-Hillel's
'operational syntax'. In place of Oswald's method of restructuring and conversion, he put forward
his categorial grammar based on the work of the Polish logician Ajdukiewicz.
At this conference, there was virtually no practical experience of programming to be
reported, or indeed of using computers. Nevertheless, their was enthusiastic talk of future
research plans. Leon Dostert suggested "the early creation of a pilot machine or of pilot
machines proving to the world not only the possibility but also the practicality of MT". In the
realisation of this pilot in January 1954, Dostert made the transition from speculation about
'mechanical translation' to the reality of 'machine translation'.
10. Further progress in 1952 and 1953
Before this historic demonstration, however, Bar-Hillel, Reifler and others published
further articles giving MT a wider audience. Reifler produced further studies on pre-editing and
post-editing, on the treatment of German compound nouns and methods for dealing with
ambiguity of homographs within sentences. Also in September 1952 James Perry at MIT
reported a simulation of word-for-word translation of a Russian text into English (Perry 1955).
The Russian words were written on separate slips of paper, drawn at random, checked in a
dictionary and English translations found, and then restored to their original order. One
simulation produced:
On/Onto/At Fig.12 traced/mapped-out/drawn parabola according-to/along/inaccord-with which move thrown/deserted with/from velocity 10m/sec. under/below
angle to/toward vertical line into/in/at 15o, 30o, 45o, 60o.
It was claimed that "the rough translations exhibited a high degree of intelligibility." To prove it,
two assistants with training in physics and chemistry wrote interpretations, e.g. for the above:
On Fig.12 a parabola is drawn according to which a body moves, thrown with the
velocity of 10m/sec and making angles of 15o, 30o, 45o, 60o with the vertical line.
Recognising possible improvements if Russian grammar were taken into account, Perry did a
follow-up simulation of a dictionary look-up procedure, which searched for the longest stem
match and then identified the ending and its grammatical function.
There continued to be reports of the SWAC computer in Los Angeles, which implied that
MT was close to realisation. For example, there was another report in the New York Times on 5th
October 1952, which was as optimistic as in 1949; and in December 1952 there was an article in
the scientific journal Industrial and Engineering Chemistry stating that "the next scientific
bottleneck to open up before the mental capacity of electronic equipment may well be the
problem of language barriers in international communication." Then it reported Perry's
experiments, quoting a simulated translation which the reporter thought was "not unmeaningful
to an individual having some knowledge of the science." A major impediment was the painfully
slow input of the documents before translation, and a photographic scanner was suggested. In all,
the report was a reasonably balanced assessment of the future potential of MT. However, the
reporter could not refrain from an amusing coda, alluding to the current Cold War witchhunting
mentality in parts of the United States:
One wonders, though, of the outcome, should one of our more ardent and less
responsible hunters of subversives get wind of these machines that may blithely
swallow the words of an alien and repugnant tongue, and then spew forth what is
possibly a strange and, therefore, suspect gibberish. Could any culpable and
innocuous computor withstand the taunts and accusations leveled, perhaps, in its
direction without manifesting an incoherent hum, suffering a twitch, or developing
at least a mild paranoia? Our machines may now also need a built-in
psychoanalytical circuit and couch for purposes of self-confession when mental
doubts becloud their electronic tubes and snarl their recording devices.
He was not to know that some 20 years later, Colby was to write a computer program PARRY,
which simulated the behaviour of a paranoid patient (Colby 1975)
During 1953, Bar-Hillel published three articles on various linguistic aspects of MT, and
two articles appeared from the UCLA researchers in the same issue of Modern Language Forum.
One was by Victor Oswald and Richard Lawson on microglossaries, the other was by Kenneth
Harper, on the importance of morphological analysis when translating Russian. In the same year
also, Anthony Oettinger from Harvard University produced a progress report on his doctorate
study for the automation of a Russian-English dictionary; his thesis was to be approved in 1954
(the first in the field of MT).
Bar-Hillel ended his two-year research post at MIT in July 1953 and returned to
Jerusalem. He was replaced by Victor Yngve, who proceeded to set up the MT research project
in the Research Laboratory of Electronics. The first progress report appeared in October. Yngve
(1955) described a simulated word-for-word translation from German into English, in which
articles and other function words were left untranslated. He reproduced an extract:
Die
CONVINCINGe
CRITIQUE
des
CLASSICALen
IDEA-OFPROBABILITY IS eine der REMARKABLEen WORKS des AUTHORs. Er HAS
BOTHen LAWe der GREATen NUMBERen ein DOUBLEes TO SHOWen: (1)
wie sie IN seinem SYSTEM TO INTERPRETen ARE, (2) THAT sie THROUGH
THISe INTERPRETATION NOT den CHARACTER von NOT-TRIVIALen
DEMONSTRABLE
PRPOSITIONen
LOSEen.
CORRESPONDS
der
EMPLOYEDen TROUBLE? I AM NOT SAFE, THAT es dem AUTHOR
SUCCEEDED IS, den FIRSTen POINT so IN CLEARNESS TO SETen, THAT
ALSO der UNEDUCATED READER WITH dem DESIRABLEen DEGREE-OFEXACTNESS INFORMS wird.
The intention was to test what could be achieved from such crude beginnings. Yngve was
surprised to find how much was comprehensible, and concluded that since "word-for-word
translations are surprisingly good, it seems reasonable to accept a word-for-word translation as a
first approximation and then see what can be done to improve it." The most obvious need was
some syntactic analysis, and this aspect of MT was to be the principal focus of research at MIT
in subsequent years.
11. The Georgetown-IBM demonstration, January 1954
The most significant outcome of the MIT conference in June 1952 was undoubtedly the
decision by Leon Dostert to start work on a computer program to demonstrate the feasibility of
MT. At Georgetown University he appointed Paul Garvin, a linguist with knowledge of Russian,
and established links with the Cuthbert Hurd and Peter Sheridan of the IBM Corporation. A
small-scale system for translating some Russian sentences into English was developed and on
7th January 1954 a demonstration took place at the New York headquarters of IBM.
The next day, the New York Times carried a front-page report of what it believed to be
"the cumulation of centuries of search by scholars for 'a mechanical translator'." Although
stressing the limited nature of the demonstration, the reporter retailed Dostert's optimism that
Such a device should be ready within three to five years, when the Georgetown
scholars believe they can complete the "literary" end of the system.
It was thought that other languages would be easier than Russian, and the dictionary data (i.e. the
punched cards) for German, French and "other Slavic, Germanic and Romance languages can be
set up at will."
Reports appeared in many other newspapers. Nearly all stressed the speed of the
operation and the fact that the punched card operator did not know a word of Russian. Spectators
were also impressed by the wide range of subject matter translated, and that the output was so
fluent.
The limitations of the experiment (a 250-word vocabulary and six grammar rules) were
freely admitted by the Georgetown researchers, but the general public had been impressed: MT
was now seen as a feasible objective, US government agencies were encouraged to support
research on a large scale for the next decade, and MT groups were established in other countries,
notably in the USSR. On the other hand, unrealistic expectations had been raised which did not
materialise for many years to come.
The IBM-Georgetown demonstration was the first actual implementation of a translation
system performing on the still primitive electronic computer. All previous work on MT had been
theoretical in the sense that none of the proposals had actually been implemented as computer
programs. The SWAC demonstrations of 1949 and 1952 had not produced translations; Huskey
and his colleagues had merely asserted that MT was possible and shown the machine on which
they believed it could be done. All other previous research in the 1950s had been either 'thoughtexperiments' or simulations on punched cards or paper slips. Before the demonstration in January
1954, the mechanisation of translation had been largely speculative. From now on, real research
began.
12. References
Bar-Hillel, Y. (1951): The present state of research on mechanical translation. American Documentation 2
(4), 229-237.
Becher, J.J. (1962): Zur mechanischen Sprachübersetzung, ein Programmierungsversuch aus
dem
Jahre 1661. J.J.Becher: Allgemeine Verschlüsselung der Sprachen (Character, pro Notitia Linguarum
Universali). Stuttgart: Kohlhammer, 1962.
Booth, A.D. and Locke, W.N. (1955): Historical introduction. In: Locke and Booth (1955), 1-14.
Colby, K. (1975): Artificial paranoia. New York: Pergamon Press.
Hutchins, W.J. (1993): The first MT patents. MT News International 5, 14-16.
Hutchins, W.J. (1994): The Georgetown-IBM demonstration, 7th January 1954. MT News International
8, 15-18.
Hutchins,W.J. (1997): From first conception to first demonstration: the nascent years of machine
translation, 1947-1954. A chronology. Machine Translation (to appear)
Locke, W.N. and Booth, A.D. (1955), eds.: Machine translation of languages: fourteen essays.
Cambridge, Mass.: M.I.T.Press.
Oswald, V.A. and Fletcher, S.L. (1951): Proposals for the mechanical resolution of German syntax
patterns. Modern Language Forum 36 (3/4), 1-24.
Perry, J.W. (1955): Translation of Russian technical literature by machine: notes on preliminary
experiments. Mechanical Translation 2 (1), 15-24.
Reifler, E. (1954): The first conference on mechanical translation. Mechanical Translation 1 (2), 23-32.
Reynolds, A.C. (1954): The conference on mechanical translation held at M.I.T., June 17-20, 1952.
Mechanical Translation 1 (3), 47-55.
Richens, R.H. and Booth, A.D. (1955): Some methods of mechanized translation. In: Locke and Booth
(1955), 24-46.
Weaver, W. (1949): Translation. Repr.in: Locke and Booth (1955), 15-23.
Yngve, V.H. (1955): Syntax and the problem of multiple meaning. In: Locke and Booth (1955), 208-266.